320-ISO 5000 K stuff

[Roberto Lequeux]

So I guess you should always expose on red using RGB histograms? And have some filters handy? What are those filters? Are they called Red Green and Blue? Somehow I imagine they won't have such simple names

I don't know but I am thinking this thread could be considered for a sticky with so many clear info jammed posts... thanks

 

[Lauri Kettunen]

What are those filters? Are they called Red Green and Blue?

In still photography one talks of color filters or cooling/warming filters.
Cooling filters are employed when the color temperature is low and they are blue. Warming filters are orange-yellowish and they are used when the color temperature is high. One may find the naming convention bit confusion, for its established in a dual sense. That is, when a yellow-orange filter is employed, then one dials a lower color temperature. In case of blue the other way around.

So I guess you should always expose on red using RGB histograms?

In my understanding the best quality is obtained using the raw view and RGB histograms. For this reason I quite can't understand why Redspace has been emphasized so much (by Red). If I've got it right, the idea of Redspace is to kind of emulate a film camera with Red One and yield a kind of safe approach giving almost what the camera can do.

And yes, if one does not YET have a good understanding of the principles can see the motivation for Redspace as an entry level approach. But, once everybody gains experience I think it is in the interest of every user to push the camera to its limits gaining the best quality. In this sense find it bit surprising how much Redspace has been emphasized.

 

[Roberto Lequeux]

Perhaps it is more geared for clients...

 

[Hans von Sonntag]

Personally, I find the RGB histo the best tool for exposure in the RedOne. Together with the RAW view (which is not RAW naturally) one gets a good clou what's going on.

The RGB histo does not only show wether things clip but also show which of the three RGB channels clip firts. Now, we can white balance this with filters. In a tungsten environment for instance the red channel will clip first. To even out the RGB levels a blueish 80C is helpful because the filter brings the 3200 K environment up to approx. 4800 K which again is pretty close to the 5000K of the sensors specs. The red channel will now clip conisderably later. By doing so we gained a bit DR, and more importantly, reduced noise in the blue channel because later in post the blue channel does not need to be pushed so much to white balance the picture.

An 81EF does the same thing in a bluish 7000K+ environment. Here, the blue channel will clip first. To create a even RGB histo an orange 81EF is doing a great job.

Several independent test showed that it is beneficial NOT to compensate colour temperature within Red's apps but later in post. It does make a lot of sense to mentain the 5.000K in RedAlert and correct the orange cast later, e.g. in a tungsten shot without an 80C, when doing the final grading. Red's apps tend to increase the blue channel's noise more than when leaving Kelvin to 5000 and correcting later in post. Obviously there are downsides to Red's ISO/Kelvin metadata concept.

Moreover, when dialing in 320 ASA and shooting a high contrast scene, in RedAlert you will find later information in highlights when dialing back from 320 ASA to 250 ASA. This is the reason why many like me find that the sensor is best represented in Red's software with 250 ASA at 5000 K. I'm not saying that Red's sensor is 250 ASA. You can rate it to what ever you want. It depends how much noise you can bare.
Red's sensor delivers a considerable amount of noise in the blacks. When underexposing to the point that skintones reach into this noise area we will get a noisy picture when trying to recover the skintone to a normal level. It's the cameraman's art to know how deep she/he can go when exposing.

What I wanted was this: Dialing in what ever is meaningful for the videovillage and still keeping my favoured metadata 250 ASA at 5000 Kelvin. Since I now use cineform and their coversion tools all the metadata mess is gone. Cineform rates the SDKs ISO at 250 ASA and I will never have to look again tediously into the shots one by another wether information in the whites is missing.

For those folks who edit the proxies in FCP and render out the timeline for finishing the metadata concept is important. But this means that the metadata has to be set on set correctly, creating a pseudo WYSIWYG workflow - for me not a viable option. I like the Red's filmlike non-WYSIWYG way of life.

Sorry for drifting a bit to the side...


Hans

Edit: Was a bit slow. Missed Lauri's post. And yes, Redspace is invented for all those who want a kind of WYSIWYG workflow with "vibrant" colours. Surely not the right path if getting the most out of the cam is the game.

 

[amourgues]

still lost...........If the camera records image at 320-ISO and the user setting is just an instruction on how to process the image, that means to me that it does not matter what iso to shoot with !...it´s only a processing path, so you can reprocess what the sensor records in post...right ?
however, if you set the iSO at, say 1000, having a lot of clipping, then you can´t recover the image by taking the ISO to 320 in post....the information is lost...it wasn´t recorded in the first place.....that would mean the sensor wasn´t recording really at 320....ahhhhhhhhh, so confused
coming back to the very basic question: what ISO to shoot with ?

 

[Eki Halkka]

The reason underexposure causes more noise and problems in post production is *not* because it causes the use of fewer levels or utilization of less of the bit depth: it's because the ratio of signal to noise (S/N) is worse.

Mainly, yes. That's why i mentioned each stop means 6dB more noise, or more accurately 6dB smaller signal-noise ratio.

But it also means digital precision is halved.

All well designed raw cameras have enough bit depth to prevent quantization at *any* exposure. Even *12 stops* underexposed.

Well, i.e. Red One is 12 bit linear. As far as i know it means that (from overexposure down):

1st f-stop has 2048 values
2nd f-stop has 1024 values
3rd f-stop has 512 values
4th f-stop has 256 values
5st f-stop has 128 values
6th f-stop has 64 values
7th f-stop has 32 values
8th f-stop has 16 values
9th f-stop has 8 values
10th f-stop has 4 values
11th f-stop has 2 values

...so, if you underexpose by 12 stops, you pretty much have everything under middle gray full black.

Of course, in practice this never happens because of noise - the sensor output is probably never zero, and most if not all of the quantizing is hidden by noise dithering.

Brighter exposure increases S/N; dimmer exposure decreases S/N. Bits are just trivia: it's the amount of light falling on the sensor that is important.

Both are important, but i.e. Red One's 12 bits are enough to capture everything the sensor is able to output above it's noise floor. If the sensor had a much better signal-noise ratio, 12 bits wouldn't be enough to store it all.

 

[Daniel Browning]

still lost...........If the camera records image at 320-ISO and the user setting is just an instruction on how to process the image, that means to me that it does not matter what iso to shoot with !

Right. The ISO setting on the camera doesn't matter. What *does* matter is how much light is coming through the lens. Your iris, shutter speed, ND, and lighting are not instructions, and cannot be changed in post production.

however, if you set the iSO at, say 1000, having a lot of clipping, then you can´t recover the image by taking the ISO to 320 in post....the information is lost...it wasn´t recorded in the first place.....that would mean the sensor wasn´t recording really at 320....ahhhhhhhhh, so confused

None of that is true. If you set the ISO at 1000, having a lot of clipping, you *can* recover the image by taking the ISO to 320 in post. The information is not lost, because it was recorded at ISO 320.

However, if you expose as if the camera was ISO 50 (which means increasing the amount of light coming through the lens by almost 3 stops), then you will lose highlight headroom, because the sensor is still at ISO 320, naively.

EDIT: I forgot to answer the last question.

coming back to the very basic question: what ISO to shoot with ?

Since the ISO setting in the camera is just part of the "instructions" or metadata, the hard part is deciding what to "rate" the camera. The easiest answer is to just follow the manufacturer's recommendation of ISO 320.

The ideal exposure is always ETTR: Expose To The Right. You can google that and figure out everything you need, but it basically works like this: decide where the very top of your highlight range is (say, a white shirt), then expose so that it is almost, but not quite, clipping. Anything brighter than that will be lost (e.g. specular highlights, reflections, etc.), but you'll maximize the dynamic range (8-11 stops). For a given scene it may result in ISO 640, 320, or even as low as 100.

 

[Daniel Browning]

That's why i mentioned each stop means 6dB more noise, or more accurately 6dB smaller signal-noise ratio.

We agree in principle, but I'm not completely with you on the exact numbers. [The highest exposure range (2048-4096) is dominated by photon shot noise and PRNU. At one stop lower, S/N is dominated only by photon shot noise. From here out, Signal (S) is decreased by 6dB, but S/N goes down much less than that (~1/3 stop), due to the fact that light falls in a Poisson distribution. After several stops, read noise begins adding to the decrease in S/N, and the drop in S/N becomes more like 1/2 stop for every one stop decrease in Signal.]

But it also means digital precision is halved.

True, but irrelevant. The drop in precision has absolutely no effect on the image whatsoever. By bringing it up, the reader may be misled to believe it is relevant or important, but it is not. What is important is the drop in S/N.

There is only one effect that precision can have on an image: quantization (in the form of posterization or quantization error).

8-bit video cameras do not have enough precision (bit depth) to record an underexposed image. So they have the negative effects of both underexposure (reduced S/N) *and* quantization. That's one reason why it's important to boost the signal in 8-bit cameras (with curves or gain). Bringing up the loss in precision is important for those cameras.

RED ONE is different because it's like a raw camera.

...most if not all of the quantizing is hidden by noise dithering.

Quantization is a complete non-issue for RED ONE. I base that on the high level of noise ever present in the last two bits and the fact that no one has posted any R3D with quantization. (Which says a lot about the quality of REDCODE, IMHO, because quantization is yet another challenge for the encoder.)

Both are important, but i.e. Red One's 12 bits are enough to capture everything the sensor is able to output above it's noise floor. If the sensor had a much better signal-noise ratio, 12 bits wouldn't be enough to store it all.

Yes; the fact that 12 bits are enough is exactly why we only have to worry about light itself (S/N) and not the bit depth.

 

[Daniel S Williams]

I don't know about you Roberto, but this thread is boggling my brain! I'm really trying to soak it in though. So many knowledgeable experts on here.

 

[Eki Halkka]

RED ONE is different because it's like a raw camera.

Red one is different because it stores the footage with high enough bit depth compared to sensor's S/N ratio. RAW bayer data versus RGB, YUV or whantnot doesn't doesn't give an advantage... quite the opposite.

Saving in linear color space is not a good thing to do precision/quantizing wise, as half of precision is lost on the brightest f-stop etc. With proper gamma curve, the distribution could be evened out, and less bits could be used - something like 10 bits should be enough to store red's sensor data in logarithmic format with similar precision as it's now saved in 12bit linear format.

...but of course, that wouldn't anymore be recording the data as sensor outputs it, so i guess it's not a good idea for a raw camera.

Quantization is a complete non-issue for RED ONE. I base that on the high level of noise ever present in the last two bits and the fact that no one has posted any R3D with quantization.

Of course quantization changes the noise characteristics a lot - in those last bits, the noise is either "on" or "off", or in other words the in-between noise values jump to the closest value.

I've never actually tested how much this actually affects the perceived image quality, but my guesstimate is that quantized noise will "look" much noisier than noise that's recorded with enough precision - as it's like adding contrast to the noise.

Anyway, i agree with your main point: that noise is more important than bit precision in red's case.

 

[Joe G.]

A lot of good answers, and yet the potential to cause a lot of confusion.

To me (please correct me if I'm wrong), the sensor is always 320asa, and RAW is recorded no matter what the camera's LUT-like ASA setting. The camera's ASA is essentially a brightness control for the viewfinder, and not a display of the accurate RAW data. The farther you try and "tweak" it from that 320 ASA point, the more potential for problems including noise or clipped highlights. Pros are used to this, by experience, and by a lot of testing. The best solution would be to test and try many configurations of exposure until it becomes second nature -- if you insist on tweaking.

The sensor and the camera's ASA setting are probably best left at 320, so as not to give a false sense of exposure safety.

As for tungsten lighting, someone (probably David Mullen) suggested tungsten for the faces, but using a more blue HMI or daylight light for background fill. The background is where the noise will probably creep in, and so you want to make sure you have enough blue light in the fill. Gelling all the lights and/or using a blue filter ("Sapphire" was recommended) like 80C, will help avoid blue channel noise.

The ETTR is highly instructive, as you want to fill the three channels with as much data as possible, so as to have that signal/noise ratio. By losing just one stop, you throw away half the potential data in that channel. Throw away more than that, and your image will suffer.

Right?

 

[amourgues]

Back to te first basic question: what ISO to shoot with ?. In simple terms, what I've got from this thread is:
- Leave ISO at 320 and temp at 5000 K
- Open the IRIS as much as you can not to clip the whitest point

did I get it right ?

 

[Daniel Browning]

Red one is different because it stores the footage with high enough bit depth compared to sensor's S/N ratio.

You're right. When I said that 'raw' is the reason that was incorrect. The real reason is as you said: sufficient bit depth. I probably made the mistake because all raw cameras have sufficient bit depth for linear recording.

Saving in linear color space is not a good thing to do precision/quantizing wise, as half of precision is lost on the brightest f-stop etc.

I wouldn't say the precision is "lost"; just wasted. Linear spends a lot of recording space on something that requires far fewer values to record accurately.

With proper gamma curve, the distribution could be evened out, and less bits could be used - something like 10 bits should be enough to store red's sensor data in logarithmic format with similar precision as it's now saved in 12bit linear format.

Agreed.

...but of course, that wouldn't anymore be recording the data as sensor outputs it, so i guess it's not a good idea for a raw camera.

REDCODE is secret sauce, so I don't know what's going on under the hood, but I would be very surprised if it was actually storing the data linearly. Nonlinear compression would take very little CPU power, is visually lossless, and would reclaim a lot of wasted space.

Of course quantization changes the noise characteristics a lot - in those last bits, the noise is either "on" or "off", or in other words the in-between noise values jump to the closest value.

I've never actually tested how much this actually affects the perceived image quality, but my guesstimate is that quantized noise will "look" much noisier than noise that's recorded with enough precision - as it's like adding contrast to the noise.

I think you're right.

 

[Hans von Sonntag]

Saving in linear color space is not a good thing to do precision/quantizing wise, as half of precision is lost on the brightest f-stop etc. With proper gamma curve, the distribution could be evened out, and less bits could be used - something like 10 bits should be enough to store red's sensor data in logarithmic format with similar precision as it's now saved in 12bit linear format.

...but of course, that wouldn't anymore be recording the data as sensor outputs it, so i guess it's not a good idea for a raw camera.

This is what CineForm is doing with the Raw signal of the SI2K. They use a 10 Bit logarithmic approach off a 12 bit AD. If I'm not mistaken Arri is going this way as well with their Arri D21. Both cameras have roughly the same DR like the RedOne, the Arri perhaps a bit more.

I guess it's a pure philosophical discussion because in most cases later in post the 12 Bit linear RAW signal will be converted into a 10 Bit RGB/YUV environment anyway. Here, logarithmic curves are doing a good job if precision is the aim, e.g. RedLog or CineForms log. conversion.

The RAW 10Bit logarithmic approach has one advantage: Less bits need to be compressed. One can use an even milder compressor for a given bitrate. Or, in Arri's case, don't use any compression - but then 10 Bit log. is a kind of compression...

Hans

 

[Hans von Sonntag]

Back to te first basic question: what ISO to shoot with ?. In simple terms, what I've got from this thread is:
- Leave ISO at 320 and temp at 5000 K
- Open the IRIS as much as you can not to clip the whitest point

did I get it right ?

As mentioned above I would set 250 ASA and 5000K for the metadata. Then in RedAlert/RC you don't have to check every single clip for lost highlights.

If you use a lightmeter rating the camera 320 ASA is a good starting point.

For exposing via the EVF/LCD I would use the RAWview mode which pictures quilte well the sensors performance. Open the iris as much as possible until white parts in your picture clip (using the RGB histogram), close the iris for 1/2 stop and you are fine.

Hans

 

[Lauri Kettunen]

Leave ISO at 320 and temp at 5000 K
- Open the IRIS as much as you can not to clip the whitest point

did I get it right ?

I would say: To maximize image quality:
- Leave ISO at 320
- Use raw view when setting exposure. Then the color temp setting has no effect on the histograms and traffic lights. Thus, the color temp setting does not matter.
- Basic trule of thumb: shoot to the right just below clipping. But if the dynamics of the target is very high --espcially if the number of stops between the bright and dark parts of the target exceeds the dynamics of the camera-- you may want to clip one of the channels. In such case avoid making the clipping are too large. A small area is typically not that distracting and thus ok.

The key point to understand is that one is not able to adjust the physical properties of the sensor.

In this sense I'm quite not able to follow Eki's argument about the bit depth. For, if I've got it right, linearity of CMOS sensors follows from the rather physical fact that the signal from a pixel of the sensor is directly proportional to the amount of photons received by that pixel. This is not adjustable, and thus half of the bits are within the last stop.

Now, of course, one may make a trade-off between bit depth and linearity by introducing a non-linear gamma map between the sensor and data storage, but then the very idea of the raw work flow is somewhat thrown away. For, linearity is the key issue of the raw work flow implying the drak, mid and bright parts of the image scale the same way in post.

This is to say, if one overexposures by two stops when shooting to the right, then in post one can come backwards the very same 2 stops as if one never had shot two stops over (to maximize the amount of data and minimize noise). But, as soon as a non-linear map, such as a log-curve, is introduced this property is lost. So, yes one can reduce the bit depth but it also has a cost. So, it's a trade-off and they are typically about a matter of taste.

 

[Roberto Lequeux]

Yes... tons of people that know more than they even need to unless they are colorists who DP or Produce their own stuff. My head is doing fine, I am wisely avoiding looking up some of the heavier quantum physics sounding stuff they are mentioning and leaving those for an other day, perhaps the weekend

What are the different ASA settings available on the camera? Is there something between 250 and 320?

And how about leaving the ISO at 320 (or slightly lower) and every so often rolling the ISO up and down to take a look "under the hood" of your file, then setting it back to where it belongs? Can you roll it fast? A wheel and real time switching would be spectacular to just quickly check areas of concern.

So what are the filters I will need? Just a warming and a cooling filter? If so are 80C and 81EF from any particular brand or a standard? Anything for green in particular?

This is to say, if one overexposures by two stops when shooting to the right, then in post one can come backwards the very same 2 stops as if one never had shot two stops over (to maximize the amount of data and minimize noise).

Here towards the end of your post, you are still talking about overexposing so long as you don't clip too much, correct? Simply getting the image nice and hot so as to take advantage of that last stop to get the most data, right?

 

[Eki Halkka]

But, as soon as a non-linear map, such as a log-curve, is introduced this property is lost.

Not really - if coded and decoded properly, overexposing for lower noise works just the same as with linear data - there's plenty of precision for each f-stop.

 

[Hans von Sonntag]

I would say: To maximize image quality:
- Leave ISO at 320
- Use raw view when setting exposure. Then the color temp setting has no effect on the histograms and traffic lights. Thus, the color temp setting does not matter.
- Basic trule of thumb: shoot to the right just below clipping. But if the dynamics of the target is very high --espcially if the number of stops between the bright and dark parts of the target exceeds the dynamics of the camera-- you may want to clip one of the channels. In such case avoid making the clipping are too large. A small area is typically not that distracting and thus ok.

The key point to understand is that one is not able to adjust the physical properties of the sensor.

Hi Lauri,

I fully agree.

What I don't understand is your proposal to mentain the 320 ASA in the metadata for maximal quality. Of course changing the metadata does not affect the recorded picture. But it does affect the "developement" of the picture afterwards in one of the Red apps. Setting metadat to 320 is following Red's advice but my experience is this: If you set metadata to 320 ASA the chance to lose some (tiny) information in the whites is quite possible unless you check every single shot and tweak exposure in RedAlert. If you set ISO at 250 ASA and 5000K you don't lose any information in the whites when using RedAlert/RC and don't need to recheck the fooatage.

I guess that Red's advice to set the camera's ISO at 320 ASA is meant for all those folks who love to view the world through RedSpace. I think be both agree that the RAWview mode is one of the best firmware features Red recently made available. RAWview is actually metadata agnostic.

Hans

 

[Hans von Sonntag]

And how about leaving the ISO at 320 (or slightly lower) and every so often rolling the ISO up and down to take a look "under the hood" of your file, then setting it back to where it belongs? Can you roll it fast? A wheel and real time switching would be spectacular to just quickly check areas of concern.

If you set ISO at 250 ASA you don't have to look under the "hood" of your file. The "hood" (love this expression) is one of the myths of RAW. There is nothing magical about RAW. Clipped whites are clipped whites. If the informatuion is not recorded there is no informartion to be found. Not even under RAW's "hood". It might give you a magical experience when you find information by setting down the ASA. But in the end it only shows that you set the metadata in the camera wrongly (for the Red apps.).

I fully understand that people set higher ASA numbers and use RedSpace because the picture looks nicer on the client's monitor. Nevertheless, one has to recheck every single shot. Pretty tedious if you ask me.

Hans